Caterpillar has spent more than three decades putting autonomous machines to work in some of the most remote and hazardous places on Earth, and it now says that experience is the template for how it deploys artificial intelligence. Speaking to TechCrunch after a fireside chat at the Ai4 conference in Las Vegas, chief technology officer Jaime Mineart laid out the through-line: the lessons from driverless haul trucks in iron-ore pits are the same lessons that decide whether a generative AI assistant or a fleet of AI agents actually earns its keep.
The story matters well beyond heavy industry. Caterpillar is a 118,000-person manufacturer that has just posted the first $20 billion quarter in its history, with power-generation sales swelling on the back of data-centre demand. It is simultaneously a customer of the AI boom, a supplier to it, and one of the few companies with a long track record of automating physical work at scale. When its CTO says the hard part of AI is not the model, people in every sector building an AI strategy should listen.
This article covers what Caterpillar automated in mining and how far that went, the specific lessons Mineart says transfer to AI deployment, the products those lessons have produced so far, the money behind the programme, and the honest limits of the mining analogy. Every figure traces to a link in the References section.
Table of contents
- What Caterpillar Learned From Three Decades of Mining Autonomy
- The Caterpillar Lesson: Deploying a Machine Is Not Transforming a Site
- Caterpillar AI Deployment in Practice: The Cat AI Assistant
- Where Else Caterpillar Is Putting AI to Work
- Caterpillar’s Numbers: Record Revenue and the Data-Centre Tailwind
- Caterpillar’s $100 Million Bet on the People Doing the Work
- What Other Organisations Can Take From the Caterpillar Playbook
- The Limits of the Mining Analogy
- Caterpillar AI Deployment FAQs
- References and Further Reading
What Caterpillar Learned From Three Decades of Mining Autonomy
Caterpillar’s autonomy push began in mining, and Mineart is clear about why: labour shortages and hazardous conditions. Remote pits struggle to staff twelve-hour shifts, and a haul truck the size of a house is one of the most dangerous machines a person can sit in. Those two pressures made mining the first industry willing to pay for driverless equipment, long before the phrase “physical AI” existed.
From one truck to a portfolio
What started with automated haul trucks has grown into a full portfolio. The company today offers autonomous drilling, autonomous underground loaders, autonomous dozers and remote-controlled construction equipment, all coordinated through a software command centre with fleet management and what the company calls remote terrain intelligence. The CES 2026 keynote described “more than 30 years of deploying autonomous machines” in mining, and the Cat MineStar Command hauling system alone has more than eleven years of operational experience behind it.
The scale of what was proven
The numbers are the reason the analogy carries weight. By October 2024, Command for hauling had moved more than 8.6 billion tonnes across hundreds of trucks at dozens of sites on three continents, spanning iron ore, copper, gold, coal, oil sands and lithium. By the following month the fleet had travelled more than 325 million kilometres. In November 2024 the company made its first autonomous deployment outside mining, at Luck Stone’s Bull Run quarry in Virginia, which passed one million tons hauled by July 2025.
| Milestone | Figure | Date | Why it matters for AI deployment |
|---|---|---|---|
| Autonomous machines in mining | 30+ years | Stated Jan 2026 | Institutional memory of what fails on real sites |
| Command for hauling in operation | 11+ years | Stated Oct 2024 | A decade of production data, not pilots |
| Material hauled autonomously | 8.6 billion tonnes | Oct 2024 | Proof the economics work at scale |
| Distance travelled autonomously | 325 million km | Nov 2024 | Edge cases encountered and engineered out |
| First deployment outside mining | Luck Stone Bull Run quarry | Nov 2024 | The transfer to dynamic sites has begun |
| Bull Run tonnage | 1 million tons | Jul 2025 | Aggregates result in under a year |
Why mining came first
Denise Johnson, group president of the Resource Industries division, has summed up the philosophy in one line: “We know it takes a combined focus on people, process and technology to ensure long-term sustainability.” That ordering — people, then process, then technology — is exactly the ordering Mineart now applies to AI, and it is the opposite of how most software-first AI projects are run.
The Caterpillar Lesson: Deploying a Machine Is Not Transforming a Site
The central insight from the TechCrunch interview is easy to state and hard to live by. Deploying an autonomous machine, Mineart argues, is not the same thing as transforming a site. A driverless truck that arrives at a mine designed for human drivers changes very little on its own; the value appears only when the dispatch process, the maintenance schedule, the shift structure and the safety rules are rebuilt around it.
“The hard part is the workflow”
Mineart put it plainly: “The hard part about autonomy and about physical AI is incorporating that technology into the customer jobsite and into the workflows.” Anyone who has watched a chatbot pilot stall after a promising demo will recognise the pattern. The technology worked; the organisation did not change around it. The answer, learned expensively in mining, is to treat change management as part of the product rather than a follow-on task for the customer.
Operators become supervisors
A second transferable lesson concerns the people already doing the job. In an autonomous mine, an experienced operator may stop driving one machine and instead oversee multiple machines from a remote command centre. The skill does not disappear; it moves up a level. The company applies the same expectation to AI: a technician using an AI assistant is still the expert, but the assistant handles retrieval and first-pass diagnosis so that expertise covers more ground.
Institutional knowledge is training data
The third lesson is subtler. The company uses its experienced operators to train the autonomy systems, capturing decades of institutional knowledge about how a site actually runs. The same principle now governs its AI work — the Cat AI Assistant is built on proprietary service data rather than on the open web, precisely because the company learned in mining that generic behaviour is not good enough for a specific pit.
| Mining autonomy lesson | How Caterpillar applies it to AI | What most AI projects do instead |
|---|---|---|
| A machine is not a transformed site | Redesign processes and roles around the AI tool | Ship the tool, hope adoption follows |
| Operators move up, not out | Technicians supervise AI-assisted diagnosis | Frame AI as headcount replacement |
| Experts train the system | Proprietary data from 1.6M connected assets | Rely on a general-purpose model alone |
| Start where risk and labour pressure are highest | Field service and legacy code first | Start with the flashiest demo |
| Prove it in production for years | Off-board assistant first, in-cab after validation | Roll out everywhere at once |
| Fund the workforce transition | $100M training pledge over five years | Leave upskilling to individuals |
Caterpillar AI Deployment in Practice: The Cat AI Assistant
The most visible product of this thinking is the Cat AI Assistant, which Caterpillar launched at CES 2026 and which Mineart told TechCrunch “is now being used by customers, operators and technicians.” It is aimed at the least glamorous and most valuable moment in the equipment lifecycle: a technician standing next to a broken machine, trying to work out what is wrong and which part to order.
Voice-first, standing next to the machine
Using voice commands, a field technician can pull up repair procedures, troubleshoot a fault and identify the parts needed before starting the repair. The company describes the assistant as a group of AI agents operating as one, drawing on the company’s Helios unified data platform. The design choice to lead with voice is itself a mining lesson: hands are busy, gloves are on, and a screen is the wrong interface for a jobsite.
Built on 1.6 million connected assets
The assistant’s real advantage is the data underneath it. Caterpillar has around 1.6 million connected assets globally and more than 16 petabytes of structured data flowing through Helios, gathered over years of telematics and service history. Chief digital officer Ogi Redzic has said this “strong digital foundation” is what lets the company “move fast and deploy new AI capabilities”. It is the industrial version of the point every IoT solutions practitioner makes: the sensors were the hard part, and the AI is the payoff.
A staged rollout, not a big bang
Consistent with the mining playbook, the rollout is deliberately phased. The off-board version of the assistant went live in the first quarter of 2026, while the in-cab version stayed in final validation, using NVIDIA Jetson Thor hardware for on-machine speech recognition and models. That mirrors how the company treated autonomous hauling — years of proof in production before expanding to a new environment.
| Layer | Component | Role in Caterpillar’s AI stack |
|---|---|---|
| Machines | 1.6 million connected assets | Telematics and service data source |
| Data platform | Helios | Manages more than 16 petabytes of data |
| Edge compute | NVIDIA Jetson Thor | On-machine speech recognition and models |
| Assistant | Cat AI Assistant | Multi-agent voice support for technicians |
| Autonomy | Cat MineStar Command | Fleet management and remote command centre |
| Engineering | Code-modernisation agents | Legacy code, test generation, defect detection |
Where Else Caterpillar Is Putting AI to Work
The assistant is the headline, but Mineart described AI threaded through the company’s operations rather than confined to one product. Three uses stand out because each maps directly onto a mining-era habit.
Site scanning and digital twins
AI now powers the company’s site-scanning software and the generation of digital twins in manufacturing. Site scanning is a natural descendant of the remote terrain intelligence that autonomous trucks rely on: a machine that has to understand a pit’s geometry to drive safely produces the same maps a planner needs. Digital twins on the factory side extend that logic to the company’s 93 manufacturing facilities, where a simulated line can be tuned before the physical one is touched.
AI agents on legacy code
The least expected use is in software engineering. “We use AI agents to modernize legacy code, generate and test new software, and identify defects earlier,” Mineart said. For a century-old manufacturer with an enormous installed base of control software, that is a serious statement, and it is consistent with the mining lesson of starting where the labour pressure is highest. Modernising legacy code is exactly the work that is hard to staff and easy to measure.
From pits to quarries and construction sites
The forward-looking ambition is to take the autonomy stack into far messier places. “Now we’re in this super exciting time where we can take all of that learning from mining and bring it into much more dynamic environments, jobsites, quarries, and construction sites,” Mineart said. Five autonomous construction machines were previewed at CES 2026, and the Bull Run quarry deployment is the first proof point. A construction site changes shape every day in a way a mine does not, which is why the company treats computer vision and on-board AI as the next frontier rather than a solved problem.
Caterpillar's Numbers: Record Revenue and the Data-Centre Tailwind
The AI conversation sits inside an unusually strong commercial moment. Caterpillar reported second-quarter 2026 sales and revenues of $20.5 billion, up 24% from $16.6 billion a year earlier and the first time the company has cleared $20 billion in a single quarter. Full-year 2025 revenue was a record $67.6 billion, and the $19.1 billion fourth quarter of 2025 had itself been a single-quarter record until Q2 2026 beat it.
Dividing $20.5 billion by the current quarter’s own figure gives the full bar below; the year-earlier quarter is $16.6 billion divided by $20.5 billion, or 81%; and the Q4 2025 record is $19.1 billion divided by $20.5 billion, or 93%.
The AI boom as a customer
Much of that growth comes from AI infrastructure rather than from AI products. Power Generation application sales reached $3.098 billion in the quarter, up 29% from $2.407 billion, “primarily in data center applications”, and the company’s earnings presentation put power-generation retail sales growth at 72% year on year. Chairman and CEO Joe Creed told analysts that on cloud and generative AI infrastructure demand, “no one is slowing down at the moment.” The backlog stood at $72 billion, with power customers placing orders out to 2030.
Setting the current quarter’s $3.098 billion as the full bar, the year-earlier $2.407 billion is 78% of it.
Why the two stories are one story
It is tempting to treat the generator business and the AI-deployment programme as separate. They are not. The company sells the engines that keep data centres running, and it uses the compute those data centres provide to run Helios and the assistant. A company that profits from the AI build-out on one side of the ledger has every incentive to prove, on the other side, that predictive analytics and generative tools deliver on a jobsite. The mining track record is what makes that proof credible.
Caterpillar's $100 Million Bet on the People Doing the Work
If there is one mining lesson Caterpillar has funded more visibly than any other, it is the workforce transition. The company employs about 118,000 people, and it has committed to spend $100 million over five years on training in AI, autonomy, robotics, digital twins and language-model skills — a pledge first announced during its centennial in April 2025 and reiterated to TechCrunch this month.
The $25 million innovation prize
Alongside the training fund, the CES 2026 keynote announced a $25 million global innovation prize over the same five-year window, and a plan to raise digital and technology investment two-and-a-half times through 2030 on top of roughly $30 billion spent on research and development over the past twenty years. The prize is $25 million against the $100 million training pledge, which is 25% — a reminder that the company is putting four times as much into its own people as into external ideas.
What the pledge signals
The point is not the size of the cheque, which is modest for a company with $11.7 billion of operating cash flow in 2025. The point is the sequencing. The company learned in mining that the operators who feared autonomy were the same people who made it work once they were trained to supervise it. Funding the training before the rollout, rather than after resistance appears, is the difference between an autonomous mine and an expensive parked truck.
What Other Organisations Can Take From the Caterpillar Playbook
Few businesses run haul trucks, but the playbook generalises. Compressed into a checklist, the Caterpillar approach to AI deployment reads like the opposite of the typical enterprise pilot.
Start where the pain is measurable
Mining was chosen because labour shortages and hazards made the return obvious. The AI equivalents inside the company are field diagnosis and legacy-code modernisation — both hard to staff, both with a clear before-and-after. Organisations building an AI strategy should hunt for the same shape of problem instead of starting with whatever a vendor demoed last.
Own the data before you rent the model
Caterpillar spent years instrumenting 1.6 million machines before it had anything for an assistant to reason over. A general-purpose model is now easy to obtain; a proprietary corpus of what actually broke and how it was fixed is not. The 16 petabytes in Helios are the asset, and the assistant is a thin layer on top of it.
Redesign the job, then deploy the tool
The lesson Mineart returns to most often is that incorporating technology into the workflow is the hard part. That means asking, before a rollout, which decisions the AI takes, which it recommends, and who supervises it — the same questions a remote command centre answers for a fleet of driverless trucks. Treat that design as the project, and the software as one deliverable within it.
Phase it, prove it, then widen it
Off-board first, in-cab after validation; mining first, then quarries, then construction. The pattern is decades long, and few companies have that patience, but the principle scales down: prove one workflow in production, measure it, and only then extend it. For a broader view of where agentic tools are heading, see our overview of AI employees and autonomous agents, or browse the latest releases in our AI models and tools hub.
The Limits of the Mining Analogy
No honest account should pretend the transfer is automatic, and Mineart himself framed dynamic sites as the next challenge rather than a finished one. Three caveats deserve space.
Mines are controlled; jobsites are not
A mine is a fenced, mapped, single-operator environment where every vehicle is on the same network. A construction site has subcontractors, pedestrians, deliveries and a layout that changes daily. The 8.6 billion tonnes of autonomous hauling were earned in the easier environment, and the aggregates deployment at Bull Run is a single site, not yet a fleet. The construction-site machines previewed at CES remain previews.
Generative AI fails differently from autonomy
An autonomous truck has a narrow, well-specified task and stops when uncertain. A voice assistant answering a technician’s question can be confidently wrong. The company mitigates that by grounding the assistant in its own data and keeping the technician as the decision-maker, but the failure mode is new, and the mining playbook offers less direct guidance on it than on workflow design.
Vendor claims still need independent proof
Most of the figures in this piece — connected assets, petabytes, tonnes hauled — are Caterpillar’s own. They are consistent across multiple releases and years, which is reassuring, but the company has not published productivity data for the assistant, and “now being used by customers” is not the same as a measured return. Readers weighing similar tools should ask for the before-and-after numbers, exactly as a mine operator would before buying a second autonomous fleet.
Caterpillar AI Deployment FAQs
What did Caterpillar actually learn from automating mining?
That deploying a machine is not the same as transforming a site. The value of autonomy appeared only when dispatch, maintenance, shift patterns and safety rules were redesigned around it, with experienced operators retrained to supervise several machines from a remote command centre. Caterpillar now applies that same people-process-technology ordering to AI.
What is the Cat AI Assistant?
A voice-driven, multi-agent assistant launched at CES 2026 that lets field technicians pull up repair procedures, troubleshoot faults and identify parts before a repair. It runs on the Helios data platform, which manages more than 16 petabytes of data from about 1.6 million connected assets, and uses NVIDIA Jetson Thor hardware for on-machine speech recognition.
How much autonomous work has Caterpillar’s equipment done?
By October 2024, Cat MineStar Command for hauling had moved more than 8.6 billion tonnes with hundreds of trucks at dozens of sites on three continents, and by November 2024 the fleet had travelled more than 325 million kilometres. The first non-mining deployment, at Luck Stone’s Bull Run quarry, passed one million tons in July 2025.
How is Caterpillar performing financially?
Second-quarter 2026 sales and revenues were $20.5 billion, up 24% year on year and the company’s first $20 billion quarter. Full-year 2025 revenue was a record $67.6 billion. Power Generation application sales rose 29% to $3.098 billion, driven primarily by data-centre demand, and the backlog stood at $72 billion.
What is Caterpillar spending on AI training?
The company has pledged $100 million over five years to train its roughly 118,000 employees in AI, autonomy, robotics, digital twins and language-model skills, alongside a $25 million five-year global innovation prize and a plan to raise digital and technology investment two-and-a-half times through 2030.
Does the Caterpillar approach apply outside heavy industry?
The mechanics differ but the sequencing transfers: pick a measurable pain point, own the data before renting a model, redesign the job before deploying the tool, and phase the rollout. The company’s own AI agents for legacy-code modernisation are a purely software example of the same method, and the same discipline underpins any serious digital transformation programme.
References and Further Reading
TechCrunch: Caterpillar is bringing to AI deployment what it learned from automating mining
PR Newswire: Caterpillar reports second-quarter 2026 results
Manufacturing Dive: Caterpillar sales surpass $20B on growing data-centre demand
PR Newswire: Caterpillar reports fourth-quarter and full-year 2025 results
Pit & Quarry: Caterpillar introduces Cat AI Assistant
IRONPROS: Cat MineStar Command enables autonomous hauling
Electrek: Caterpillar autonomous haul trucks reach one-million-ton milestone
Manufacturing Dive via Yahoo Finance: Caterpillar pledges $100M to upskill workforce
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.